A Microphysiological HHT-on-a-Chip Platform Recapitulates Patient Vascular Lesions

Hereditary Hemorrhagic Telangiectasia (HHT) is a rare congenital disease in which fragile vascular malformations (VM) – including small telangiectasias and large arteriovenous malformations (AVMs) – focally develop in multiple organs. There are few treatment options and no cure for HHT. Most HHT patients are heterozygous for loss-of-function mutations affecting Endoglin (ENG) or Alk1 (ACVRL1); however, why loss of these genes manifests as VMs remains poorly understood. To complement ongoing work in animal models, we have developed a fully human, cell-based microphysiological model based on our Vascularized Micro-organ (VMO) platform (the HHT-VMO) that recapitulates HHT patient VMs. Using inducible ACVRL1 -knockdown, we control timing and extent of endogenous Alk1 expression in primary human endothelial cells (EC). Resulting HHT-VMO VMs develop over several days. Interestingly, in chimera experiments AVM-like lesions can be comprised of both Alk1-intact and Alk1-deficient EC, suggesting possible cell non-autonomous effects. Single cell RNA sequencing data are consistent with microvessel pruning/regression as contributing to AVM formation, while loss of PDGFB implicates mural cell recruitment. Finally, lesion formation is blocked by the VEGFR inhibitor pazopanib, mirroring positive effects of this drug in patients. In summary, we have developed a novel HHT-on-a-chip model that faithfully reproduces HHT patient lesions and that can be used to better understand HHT disease biology and identify potential new HHT drugs.

Additional Declarations: Yes there is potential Competing Interest.CCWH and APL are co-founders of, and have an equity interest in, Aracari Biosciences, Inc., which is commercializing the vascularized microtumor model.All work is with the full knowledge and approval of the UCI Con ict of Interest Oversight Committee.None of the other authors have any potential con icts of interest to disclose.

Introduction
Hereditary Hemorrhagic Telangiectasia (HHT) is a rare congenital disease that impacts 1 in 5000 people and is characterized by the sporadic development of vascular malformations. 1,2These disorganized lesions include small, dilated, and fragile tangles of vessels called telangiectasias 3 that affect skin and mucosa, especially in the nose, where their rupture leads to frequent and uncontrolled epistaxis (nosebleeds) that greatly impacts patient quality-of-life. 4,5Many HHT patients also develop larger arteriovenous malformations (AVM) -enlarged arteriolar-to-venous shunts 6 -within their gut, lung, liver, and brain that can compromise tissue perfusion leading to organ damage and eventual failure. 4HHT is sub-divided into disease subtypes based on patient gene mutation status.Greater than 90% of HHT patients inherit loss-of-function mutations affecting either ENG (HHT1), which encodes Endoglin, or ACVRL1 (HHT2), which encodes Activin-like Kinase receptor 1 (Alk1). 4Eng and Alk1 are endotheliumexpressed cell surface TGF-β superfamily co-receptors, with Alk1 being the signal transducing component, and mutation of either affects Alk1-driven signaling through downstream Smads.Speci cally, Alk1 receptor activation phospho-activates the Smad1/5 complex, which then binds to Smad4.The activated Smad complex translocates into the nucleus to regulate target gene expression. 7rare variant of HHT -juvenile polyposis-HHT (JP-HHT) 8 -accounts for ~ 2% of HHT cases and has been mapped to the SMAD4 gene, a key integrator in the Smad pathway.[8][9][10] Importantly, patients are heterozygous for Alk1, Eng, or Smad4 mutations, with lesions being thought to occur as a result of somatic mutation of the residual allele.[11][12][13] A recent study that used deep sequencing of isolated HHT patient lesions has con rmed this hypothesis, but with the surprising additional observation that many of the lesions in the same patient are of mixed composition, including both heterozygous and subpopulations of homozygously-mutated endothelial cells (EC) containing different second hitssuggesting multiple loss-of-heterozygosity (LoH) events. The relevant ligands for the Alk1/Eng receptor complex are BMP9 and BMP10, 15 although other TGF-β ligands are also capable of binding at much higher concentrations.15 Of these, BMP9 is present in blood at concentrations well above its binding a nity for the Alk1 receptor, suggesting that there is tonic signaling through this pathway in established vasculature, possibly to maintain EC quiescence.16,17 Unfortunately, there is currently no cure for HHT and few treatment options, due in part to the challenges of modeling HHT vascular malformations in the preclinical setting.Most research on the pathogenesis of vascular malformations has been conducted in genetically-modi ed mouse models wherein Alk1, Eng, or Smad4 are homozygously ablated, either globally or only in vascular endothelium.18 Although these models have provided important insights into HHT disease biology, 9,10,14,[19][20][21] AVMs occur only sporadically and can be di cult to identify and monitor in real-time. Seveal key insights have also been provided by zebra sh models, including the insight that Alk1 is a ow-sensitive gene and that dysregulated ow-directed EC migration may contribute to vascular malformation in HHT.[22][23][24] In this study, we use a microphysiological system (MPS) platform to generate a human cell-based in vitro microphysiological model of vascular malformations in HHT that can complement and enhance research in animal models.
MPS --or Organ-on-a-chip (OoaC) --refers to the engineering of arti cial microtissues in a threedimensional in vitro microenvironment that captures the microphysiology of native, intact tissue. 25,26PS models enable reproducible and high-throughput engineering of complex and physiological microtissues in an in vitro setting.They are thus powerful new tools for understanding healthy and diseased tissues that will help in translational research by supporting both compound library drug screens and personalized medicine applications. 25,278][29][30] Under these conditions, EC self-organize over the course of 4-5 days into a lumenized microvascular network that is perfused by a gravity-driven blood substitute.We have previously used the VMO platform to study vascularized micro-tumors, 28,[31][32][33] as well as to model healthy tissue-vascular interactions. 29,34In this study, we describe an adapted VMO platform that allows for the development of disorganized vessel lesions with the architectural hallmarks of HHT telangiectasias and AVMs.We call this novel HHT-on-a-chip microphysiological platform the HHT-VMO and show that it can be used to provide valuable insights into HHT disease biology and to evaluate drugs for their therapeutic potential in HHT.

Results
Alk1-de cient EC form hyperdense microvasculature in vitro.
As noted above, HHT patients are germline heterozygous for loss-of-function mutations that affect either Alk1 (ACVRL1) or its co-receptor Eng (ENG). 4To generate EC lacking intact Alk1 signaling, mimicking the LoH events thought to occur in lesions, we initially pursued a transient silencing RNA (siRNA) approach to target endogenous ACVRL1 (si-ACVRL1) transcript.This achieved a near-complete knockdown of Alk1 mRNA that persisted for at least 7 days in primary human umbilical vein EC (HUVEC), but this knockdown fully recovered to control levels by day 12 (Supplemental Fig. 1A).We seeded EC transfected with si-ACVRL1 (or scrambled si-Ctrl) into our previously described VMO platform 28,29,31,35 at equal starting cell densities (Supplemental Fig. 1B-C) and observed that microvasculature that formed from si-ACVRL1 EC were hyperdense relative to si-Ctrl (Supplemental Fig. 1D-I).We observed a similar phenotype with siRNA knockdown of ENG (Supplemental Fig. 2), suggesting that this hyperdense phenotype arises from loss of signaling through the Alk1/Eng HHT-associated signaling pathway.
Next, to generate stable Alk1 knockdown that persists for longer than 7 days, we used lentiviral transduction to introduce an Isopropyl ß-D-1-thiogalactopyranoside (IPTG)-sensitive ACVRL1-targeting short hairpin RNA (shRNA) into primary HUVEC at > 85% transduction e ciency (Supplemental Fig. 3A-B).IPTG is an allolactose mimetic frequently used as an inducing agent to activate tunable transcription of genes under control of the lac operon, such as in our IPTG-inducible sh-ACVRL1 system.In resulting sh-ACVRL1 EC, addition of IPTG into the perfusion medium induced dose-dependent Alk1-de ciency (Supplemental Fig. 3C).3mM IPTG produced > 90% ACVRL1 mRNA knockdown, and transduction of the construct into EC was stable as evidenced by a persistently high-e ciency of ACVRL1 gene knockdown at day 14 in response to IPTG (Fig. 1A).Importantly, we tested two independent shRNA sequences (Table 1) and results were indistinguishable (data not shown).Consequently, we used them interchangeably in the reported studies.Alk1 protein expression was also reduced in sh-ACVRL1 EC in response to IPTG (Fig. 1B).Consistent with our initial experiments using siRNA (i.e.short-term Alk1 or Eng knockdown, Supplemental Figs.1-2), IPTG-treated (i.e.Alk1-de cient) sh-ACVRL1 EC formed hyperdense microvasculature in our published VMO 28,32 platform (schematic in Supplemental Fig. 4) compared to untreated controls (Fig. 1C).This observed increase in microvessel density was quanti ed in FIJI/ImageJ as a signi cantly increased vessel area (Fig. 1D), increased total vessel length (Fig. 1E), increased branchpoint number (Fig. 1F), and a decreased mean lacunarity (Fig. 1G).At later times overall vessel length was similar in control and knockdown, likely due to the physical constraints of the system.Perhaps related to this, vessel diameter was increased at later times (Fig. 1H).Importantly, IPTG alone had no effect on microvascular network formation in the VMO (Supplemental Fig. 5A-D).Interestingly, even though knockdown of Alk1 produced higher microvessel density and more enlarged vessels in the VMO, vessel network architecture remained relatively ordered in appearance, with no obvious development of disorganized focal lesions that resembled AVMs.
HHT-VMO platform supports physiological ow at arteriolar levels.
In mouse models of HHT, vascular malformations appear in the neonatal retina in regions associated with higher uid shear stress, whereas at lower levels of shear, vascular overgrowth (telangiectasia-like) is seen, 36 suggesting that hemodynamic ow is a critical initiating signal for the development of these lesions.Using in silico approaches, we modeled ow in our VMO platform and found that a maximal uid velocity of 2µm/s, producing a maximal shear stress of < 0.5 dynes/cm 2 (Supplemental Fig. 6).These values are signi cantly lower than the ow velocity and shear stress values calculated for the regions predisposed to shunt formation in the retinal vasculature of neonatal mice. 37 generate a platform with a more physiologic range of intravascular ow and uid shear stress, we adapted our base VMO (Supplemental Fig. 4) by uncoupling the upper and lower micro uidic channels and enlarging the central tissue chamber and media reservoirs (Fig. 2A).This indeed led to higher levels of interstitial ow (Fig. 2B), but still in the range necessary to induce formation of a perfused microvascular network.Network formation took 6-8 days following initial cell seeding (Fig. 2C), consistent with how the microvasculature develops in our published VMO models. 28,29,31,35Once vessel formation was complete, addition of 70kDa uorescent dextran to the media reservoirs marked lumenized microvessels and revealed that vessels are perfused and that ow is only through vessels (Fig. 2D).To model the normal developmental process, we next developed a two-step ow protocol to support the development of a perfused microvasculature under physiologic ow, wherein EC are initially exposed to bidirectional low ow for the rst 6 days, mirroring the hemodynamic conditions of developing microvessels.After 6-8 days, networks were switched to unidirectional high (physiological) ow, consistent with the conditions present upon establishment of a systemic circulation in vivo (Fig. 2E).In silico modeling of low-(Fig.2F) and high-ow (Fig. 2G) conditions con rms that they support physiologic uid shear stress levels of ~ 1 dynes/cm 2 and ~ 5 dynes/cm 2 , respectively.Importantly, we observed increased expression of the endothelial-speci c arterial marker Ephrin B2 (EFNB2) only in regions of high shear stress following 2 days of high ow in the HHT-VMO platform (Fig. 2H).In parallel, perivascular cells in the high shear stress region of the microvasculature also acquire expression of smooth muscle α-Actin (SMA) (Supplemental Fig. 7), another indicator of vessel arterialization under these conditions.Alk1-de cient EC in the modi ed platform generate disorganized microvascular networks with lesions reminiscent of HHT patient telangiectasias.
We generated EC transduced with the IPTG-sensitive sh-ACVRL1 construct, which were then seeded into the HHT-VMO platform.IPTG was added to the medium beginning at day 0 and maintained throughout the timecourse of the experiment to induce knockdown of endogenous Alk1 expression.Wild-type (control) networks were not exposed to IPTG, thereby leaving Alk1 expression intact.Consistent with our observations in the original (low-ow) VMO platform (Fig. 1), IPTG-treated (Alk1-de cient) EC in the modi ed platform also formed hyperdense networks compared to wild-type controls when the shear was at the lower end of physiologic levels (Fig. 3A).This was quanti ed as a signi cantly increased vessel area, total vessel length and branchpoint number, as well as signi cantly decreased mean lacunarity, by day 6 (Fig. 3B-E).When networks were switched to high ow in the presence of IPTG, we observed microvascular structures that progressively grew over time (Fig. 3F), resulting in enlarged and dilated focal lesions that varied in their presentation from dense clusters or tangles to enlarged loop-like structures (Fig. 3G).This disorganized architecture is reminiscent of HHT patient telangiectasias, which are also highly variable and typically present in the nasal septum as "spots", "loops", "spiders", or "raspberry"-like tangles of disorganized microvasculature. 3nsistent with the observed appearance of dilated telangiectasia-like lesions with addition of circulating IPTG, morphometric analysis of the entire microvascular network revealed a signi cant increase in the frequency of vessels with a diameter > 50 µm and a concomitant decrease in the frequency of vessels with a diameter < 50 µm (Fig. 3H).Using an investigator-blinded semi-quantitative scoring approach (see Materials and Methods), we also detected a signi cant increase in vessel disorganization score in IPTG-treated networks (Fig. 3I).Importantly, there was no effect on vessel network appearance when wild-type EC (Supplemental Fig. 5E-H) were treated with IPTG in the platform.
Lesion formation only occurs when Alk1 is lost early in vessel development.
To assess when Alk1 expression is most critical during vessel development, we seeded sh-ACVRL1 EC into the HHT-VMO and applied IPTG to the perfusion medium prior to network formation (day 0, as described above), during active network formation (day 3), or after the vasculature was largely established (day 8).We then assessed network morphology at days 10 and 14, to allow a minimum of 6 days in each group for lesions to develop after onset of IPTG; these timepoints were selected based on our earlier studies (Fig. 3) showing that lesions became obvious by day 6 when IPTG was administered at day 0. Remarkably, only the administration of IPTG at day 0 or day 3, but not at day 8, led to the appearance of disorganized vascular lesions (Fig. 4).This indicates that there is a critical early window during vessel network development when the absence of Alk1 can trigger lesion formation.In the HHT-VMO system, this window is prior to 8 days.These ndings are consistent with work from the Oh lab 14,19,20 , which has used mouse models of HHT to show that loss of Alk1 expression drives lesion formation in the early phases of an angiogenic response, whereas loss of Alk1 in stable vessels in the absence (or neutralization) of pro-angiogenic signaling does not produce a phenotype.

Pazopanib prevents lesion formation.
Pazopanib is a VEGFR inhibitor that has been shown to improve bleeding symptoms in HHT patients, 38,39 although the drug's effects on underlying vascular malformations in HHT patients remains uncertain.To assess whether HHT-VMO lesions could be prevented by application of pazopanib, we seeded shACVRL1 EC into the HHT-VMO and added IPTG from day 0. We then added 25nM (data not shown) or 100nM pazopanib (or DMSO vehicle) to the perfusing medium beginning at day 3, when our data indicate that Alk1 remains necessary to prevent vascular lesions (Fig. 4).Consistent with the clinical data showing bleeding and anemia improvement with clinical administration of pazopanib, 38,39 we observed an inhibition of lesion formation when pazopanib was added to IPTG-treated networks (Fig. 5) with resultant vascular networks being barely distinguishable from wild-type networks.
Enrichment for strongly Alk1-de cient EC drives lesions reminiscent of arteriovenous malformations.
The approach described above reproducibly produced disorganized and dilated lesions reminiscent of HHT patient telangiectasias, but shunt-like structures did not reliably develop.We hypothesized that this might be because Alk1-de ciency is heterogeneous across the transduced EC population due to the relatively low MOI used (see Materials and Methods) along with a normal distribution of lentiviral uptake; thus, there might be incomplete penetrance of Alk1 knockdown.To remedy this situation and to enrich for more consistency in Alk1-knockdown, we used a brief puromycin pre-selection prior to cell seeding in the HHT-VMO (or in a modi ed version of the platform redesigned to t a standard 96-well plate format while keeping all other parameters consistent, Supplemental Fig. 8).To mimic the in vivo situation, where only some EC are homozygous mutant due to an LoH event, 12 we co-seeded wild-type EC along with puri ed sh-ACVRL1 EC in a one-to-one ratio.Remarkably, we now observed that on addition of IPTG, we could consistently develop shunt-like structures (Fig. 6).Conduits were especially obvious when intravascular ow was marked by perfusion of 70kDa uorescent dextran, which owed directly from the upper (arteriole-like) to the lower (venule-like) micro uidic channels, through these enlarged vascular conduits (Fig. 6A).
A key feature of a developing AVM is that due to its increasing diameter, it carries a greater fraction of the available blood, and thereby starves adjacent branching microvessels.These branching collateral microvessels eventually regress, leaving an enlarged direct artery-to-vein connection -an AVM -in place of the previous balanced capillary bed.We observed a similar phenomenon in the HHT-VMO, and again, this was particularly apparent when vessels were perfused with uorescently-labeled dextran.Shunt-like vessels were strongly uorescent, whereas the adjacent capillary-like vessels became extremely narrow and appeared to cease supporting ow (Fig. 6B).Disorganized, shunt-like lesions formed in 63.2% of IPTG-treated networks, compared to only 11.1% of control networks (Fig. 6C), and these vascular structures typically formed by day 8, two days after networks were switched from bidirectional low ow to unidirectional high ow (Fig. 6D).In the retina of Alk1 mutant mice, shunts develop preferentially in rst-order retinal vessels -i.e., vessels that branch immediately from the primary arteries that feed retina -that are exposed to high uid shear. 36Consistent with this nding, we also observed that shunts in the HHT-VMO appeared more reliably in rst-order vessels in HHT-VMO networks -i.e., the rst branching vessels immediately downstream from the upper micro uidic channel where uid velocity and shear stress is calculated to be highest (Fig. 2F-G).These shunt-prone rst-order vessels were measured at day 10 for their vessel diameter at the upper, middle, and lower regions of the tissue chamber.In networks containing Alk1-de cient EC, the mean diameter of shunt-like structures was 75.0 ± 19.1 µm, compared to 46.6 ± 11.7 µm in Alk1-intact (i.e., no IPTG) control networks (Fig. 6E).This corresponds to a 61% increase in rst-order vessel diameter in Alk1-de cient vessel networks compared to wild-type controls, and a 20-fold increase in ow, consistent with the development of a shunt.
To better understand the progression by which rst order vessels enlarge into shunt-like conduits in Alk1de cient networks, we imaged uorescent vessels to track shunt formation over time.At day 4, overall microvascular architecture appears relatively organized, although focal, disorganized telangiectasia-like structures -if not shunts -are already apparent.This is consistent with the early hyperdense phenotype we reported at this timepoint in Fig. 3. Interestingly, when we used in silico modeling to understand how intravascular ow is affected by the presence of these telangiectasia-like lesions, we found that ow (and thus uid shear stress) was variable throughout the length of these rst-order vessels (Fig. 6F, inset) Speci cally, we found that shear is elevated immediately downstream of a dilated, telangiectasialike tangle (Fig. 6F, day 4 inset).By day 10, we found that these high-shear regions had outwardly remodeled to increase vessel diameter resulting in reduced and more homogenous shear -a normal vascular response to high focal ow and shear to redistribute ow primarily through these large shuntlike conduits -resulting in an apparent AVM (Fig. 6F, inset).Further studies are needed to con rm this nding at greater time-and uorescent-imaging resolution.
Shunts are comprised of both Alk1-intact and Alk1-de cient EC.
A reasonable hypothesis regarding lesion formation in HHT is that a single clone of HHT-mutant EC expands to dominate a region of the vascular network, and that this region undergoes AVM formation.
This has been previously observed to drive vascular malformation in diseases such as Cerebral Cavernous Malformation (or, CCM). 40To assess this hypothesis, we expressed red or green uorescent reporters in wild-type (Alk1-intact) and sh-ACVRL1 (Alk1-de cient) EC populations, respectively, in the HHT-VMO.We applied IPTG (or not) beginning at day 0 to induce Alk1 knockdown selectively in sh-ACVRL1 EC.For this study, because only rhodamine-or FITC-conjugated dextrans were available, EC were imaged on red and green channels rst, and then immediately re-imaged following administration of 70kDa rhodamine dextran to perfusing media to mark intravascular ow.Pre-and post-dextran images were then stitched together, manually overlaid, and differentially false-colored in order to distinguish each (Alk1-intact vs. Alk1-de cient) EC population from themselves and from the perfused uorescent dextran.In control networks (lacking IPTG) wherein both EC populations (blue and yellow, Fig. 6G) express Alk1, EC were integrated evenly into the network at day 8 and both telangiectasia-like and AVM-like vascular lesions did not develop.Instead, all microvessels were well perfused as determined by distribution of circulating 70kDa uorescent dextran (Fig. 6G).Perhaps surprisingly, and certainly contrary to our initial hypothesis, on addition of IPTG to knock down Alk1 in half of the cells (Alk1-de cient, yellow), we found that AVMs formed and that these were chimeric, compromised of an intermingling of blue (Alk1-intact, wild-type) and yellow (Alk1-knockdown) EC (Fig. 6G and inset).This has profound implications for the mechanism of shunt formation, as discussed below.Importantly, no aberrant vessel morphology was observed when EC transduced with control vector were co-mixed with wild-type EC (Supplemental Fig. 9).
Gene expression changes in Alk1-de cient microvasculature.
Consistent with these data, gene ontology (GO) analysis of upregulated genes following Alk1 knockdown (EC2) showed an association with EC migration (Fig. 7H).In contrast, GO analysis of upregulated genes in control networks showed an association with establishment of endothelial cell barrier, regulation of angiogenesis, and regulation of vasculature development (Fig. 7H).Finally, pathway analysis focusing speci cally on the PDGF pathway highlighted the loss of PDGFB-PDGFRβ signaling from EC to pericytes and SMC (Fig. 7I).Thus, EC2 appears to represent a population of cells that are actively remodeling vessels from a stable and complex vascular network to a relatively small number of large AVM-like vessels.

Discussion
Owing to their low-cost, reproducibility, and ease-of-use, OoaC/MPS technologies offer new avenues for the study of healthy physiology, disease biology, and preclinical drug screening using fully human tissues. 25,45,46In particular, vessel-on-a-chip platforms are a potentially powerful tool for studying the pathogenesis of vascular malformation, especially in designs where EC are not constrained to form vessels along pre-established micro uidic channels and structures.8][29][30] Recently, Soon and colleagues 47 used a similar design to model brain arteriovenous malformations (AVM) by comixing in EC expressing mutant variants of KRAS4.Resulting networks were abnormal in appearance and included regions of dilated vessels.This study underscores that vessel-on-a-chip technology that incorporates ow can serve as a basis for modeling vascular malformation defects relevant to several disease contexts, including HHT.Indeed, in the present work, we were able to show a hyperdense and abnormal phenotype in the HHT-VMO using both siRNA and shRNA to knockdown Alk1 in primary human EC.However, shunt-like structures did not develop in the base VMO platform, suggesting that even in the presence of Alk1-de ciency, additional signals are required to drive shunt formation.
Recently, Orlova and colleagues seeded HHT1 patient-derived human iPSC-derived (ENGhaploinsu cient) EC (iPSC-EC) into a vessel-on-a-chip platform.Resulting networks were distinct in morphology compared to microvessels formed from isogenic control iPSC-EC; surprisingly however, although microvessels were more permeable and fragile, HHT-mutant microvascular networks had reduced (rather than increased) mean vessel diameter and vascular density, suggesting that HHT-like vascular lesions did not develop in this setting.One possible explanation is that the signals that drive HHT-associated vessel malformations (i.e., telangiectasias and shunts) may not be su cient in some vessel-on-a-chip designs to produce de nitive vascular malformations.For example, patient-derived iPSC-EC are heterozygous for HHT-causing gene mutations and retain residual gene expression from their remaining intact allele; as such, a second-hit LoH that ablates residual Alk1 or Eng gene expression -a proposed precursor to lesion formation in HHT patients 11,12 -may be absent.
We and others 22,36,48 believe that an additional critical signal for AVM formation in HHT is hemodynamic ow.Using an HHT mouse model, Larrivee and colleagues showed that shunts typically form in high-ow regions of the neonatal mouse retina close to the optic nerve, 36 where in silico modeling predicts wall shear stress values are higher than in the retinal vascular periphery, 37 and where the retinal vasculature typically undergoes EC quiescence, arteriovenous speci cation, and vessel organization and maturation. 49,50Furthermore, Rochon and colleagues show in zebra sh models that ow-sensitive EC migration is dysregulated with Alk1-de ciency 22 which may re ect a defect in mechanosensitive signaling transduction in mutant EC.In the current study, we nd that gravity-driven ow in our base VMO platform produces shear stress values of ~ 0.5 dynes/cm 2 --clearly su cient to support the formation of vessels-on-a-chip, 28 but apparently insu cient to support the development of AVMs (as evidenced by our nding that Alk1-de cient EC do not generate shunts in the base VMO, Fig. 1 and Supplemental Fig. 1).
To address these and other issues, we redesigned our existing VMO platform to support a signi cantly increased gravity-driven intravascular ow, resulting in ten-fold greater wall shear stress values of ~ 5 dynes/cm 2 .We also enlarged the tissue chamber to support a broader range of ow and shear stress pro les across the resulting microvascular network.The HHT-VMO design supports microvascular network formation similarly to the original VMO, and higher ow seems to support additional arteriovenous speci cation.Seeding of Alk1-de cient EC and growth under lower ow conditions produces hyperdense microvessel networks with development of enlarged and dilated focal lesions reminiscent of HHT patient telangiectasias.Further enrichment for more penetrant Alk1-de ciecy in EC drove the development of shunt-like AVM structures under higher ow conditions.Importantly, we demonstrate in several experiments that lesions respond similarly to in vivo HHT mouse models.For example, consistent with published work in mice, 19,20 Alk1-de ciency during periods of active vessel growth produces lesions in the HHT-VMO, whereas loss in established vessels does not.We also showed that treatment with the VEGFR inhibitor pazopanib -which appears to reduce HHT-associated bleeding in patients -prevented the appearance of lesions in Alk1-de cient networks.
An important nding in this study is that lesions are comprised of both wild-type (i.e., Alk1-intact) and Alk1-de cient cells, and that these are seemingly randomly inter-mixed.This is in contrast to a model, with some appeal, whereby a clone of mutant cells expands to dominate the vessel wall, triggering lesion formation, as is reported to occur in Cerebral Cavernous Malformation. 40Instead, these data favor a cell non-autonomous mechanism for shunt formation and are consistent with the recent work of Snellings and colleagues 12 which reported that HHT patient lesions are heterogeneous for both HHT geneexpressing EC and gene-ablated EC (by a proposed second-hit LoH event).These data are also consistent with mouse studies showing that minimal Endoglin deletion in a subpopulation of EC is su cient to produce a lesion when combined with a pro-angiogenic stimulus. 14It is not yet clear how the presence of randomly positioned mutant EC can trigger a vessel-wide lesion, however our gene expression data hint that cell-cell communication may well play a major role as we noted mis-regulation of several junctional and signaling molecules in our single-cell RNA sequencing analysis (Fig. 7).Further studies are needed to address these ideas, including by taking advantage of the HHT-VMO's ability to support real-time observations of lesion formation (as presented in Figs. 3 and 6).We also see evidence in the scRNA-seq data for a population of EC (EC2) that appear to be associated with vascular remodeling.During shunt formation in the HHT-VMO we see regression of smaller vessels concomitant with the growth of the shunt, which would be consistent with integration of the regressing cells into the larger vessel.Several genes known to be involved in vessel stability or pruning were differentially expressed in EC2 compared to other EC, including KLF2, FZD4, NOS3, TSPAN12,TAGLN2, CDKN1A, and MAPK10.Interestingly, PDFGB, which is a driver of mural cell recruitment to the vessel wall, was also reduced in knockdown cells leading to a loss of PDGFB-PDGFRβ (Fig. 7I).Lebrin et al 51 have previously shown that thalidomide treatment restores mural cell coverage in an animal model of HHT, at least in part through upregulation of PDGFB.3][54] However, whether there was a loss of PDGFB expression in the mutant mice was not reported.Finally, the expression of multiple matrix and matrix remodeling genes, along with expression of both pro-and anti-angiogenic semaphorins, marks the EC2 cluster as a unique population of EC, driven by loss of BMP9-Alk1-mediated regulation, that seems to play a critical role in the dissolution of the vascular network and the formation of AVMs.
In summary, the HHT-VMO is an HHT-on-a-chip platform that reliably produces vascular lesions with morphological similarity to the telangiectasias and AVMs that arise in HHT patients.Importantly, our ndings match those seen both in mouse models 9,18,20 and in patient studies. 3,6In addition, we have used the model to probe both the timing and the mechanism of lesion formation, and we nd that AVM generation is a cell non-autonomous process.Thus, the HHT-VMO captures much of the physiology of intact tissue while offering the increased experimental reproducibility, scale, and ease-of-use bene ts of in vitro models.Based on this work, we further adapted the HHT-VMO device for a standard microwell plate format, which allows us to integrate readouts that rely on equipment that are tted for devices with a microplate footprint.Thus, we have signi cantly increased the scale of studies that can be supported by the HHT-VMO; importantly, mid-throughput drug discovery screens are certainly feasible.We therefore believe that the HHT-VMO can complement and augment existing in vivo models of HHT, and that together these tools will further propel research into HHT and the pathogenesis of vascular malformation.

Methods
Cells.HUVEC were selected as our generic EC type for this study because they are easy to isolate and culture, because they retain EC identity for multiplate passages, and because we nd that they reliably produce highly organized vascular networks in VMO platforms (e.g., Figs.1-2) with a range of vessel calibers within the microvascular range (data not shown).We have previously also shown that HUVEC do not endogenously express high levels of arteriovenous speci cation markers, but are capable of acquiring expression of these speci cation markers in response to physiological ow. 49For this study, primary human umbilical vein endothelial cells (HUVEC) were isolated as previously described, 55 and expanded in monolayer culture on gelatinized tissue culture asks in either M199 medium (Thermo #11150067) supplemented with 10% fetal bovine serum (FBS) (Gibco #160000044), 50 µg/mL Endothelial Cell Growth Supplement (ECGS), (Corning #354006) and 50 µg/mL gentamicin (Thermo #15710064); or, in complete EGM-2 (Lonza CC-3162).Primary normal human lung broblasts (NHLF, Lonza CC-2512) were cultured in Dulbecco's Modi ed Eagle Medium (Corning #10-017-CV) supplemented with 10% FBS (Gicbo) and 50 µg/mL gentamicin (Thermo).All cell lines were routinely tested for mycoplasma contamination, and primary cells were discarded after 8 passages in culture.siRNA.HUVEC were transfected with pooled si-ACVRL1 (Ambion #4392420) or si-ENG (Dharmacon #L-011026-00) or non-targeting scrambled control (Dharmacon #D-001810-10 or Ambion #4390843) constructs, as previously described. 49In brief, EC were cultured in complete EGM2 over-night prior to transfection (via Lipofectamine 2000, Thermo #11668019) with 8nM siRNA for 24 hours.EC were then washed with fresh complete EGM2 for an additional 24h-48h prior to additional experiments.
Fluorescent Reporter and shRNA Lentivirus.All lentivirus in this study was packaged with generation II lentiviral packaging vectors in 293T cells, and lentiviral-containing supernatant was tested for mycoplasma contamination and then concentrated in 50% polyethylene glycol (PEG) 8000 (Promega V3011) for at least 72h at 4°C.All lentiviral batches were then tested using HUVEC to calculate concentration of active lentiviral particles.For uorescent reporter expression, HUVEC were transduced with lentivirus containing CMV promoter-driven red (mCherry, LegoC2), green (eGFP), or blue (Azurite) uorescent reporters using 8µg/mL polybrene (Fisher # NC9840454) at a multiplicity of infection (MOI) ≈ 2-5.This typically yields 80-90% transduction e ciency.To achieve inducible Alk1 knockdown, individual ACVRL1-targeting short hairpin-RNA (shRNA) clones (Table 1) were inserted into the MISSION® 3X-LacO Inducible shRNA plasmid backbone (Sigma).Both clones target a similar region in the ACVRL1 transcript and produced similar levels of knockdown; thus, both were used interchangeably for this project.sh-ACVRL1 was transduced into uorescent EC at an MOI ≈ 0.8-1 in the presence of 8µg/mL polybrene.Using a higher MOI compromised network formation in the VMO or HHT-VMO device, even in the absence of IPTG.Following transduction, EC were recovered from lentiviral transduction with fresh complete EGM2 for a minimum of 24h, at which point they were seeded immediately into the VMO or HHT-VMO; or, for some studies, EC were enriched for high sh-ACVRL1 transduction by being cultured for 48h in complete EGM2 containing 1µg/mL puromycin (Sigma #8833) and recovered in fresh complete EGM2 for an additional 48h prior to co-seeding into the HHT-VMO device.To knockdown endogenous Alk1 expression, sh-ACVRL1 expression was induced by addition of up to 3mM IPTG (Sigma #16758) in static monoculture, or in circulating media of VMO or HHT-VMO devices.qPCR.Cells in monoculture were washed in PBS and ash-frozen at -80°C or immediately processed for RNA.RNA was isolated in TRIZol™ reagent (Invitrogen) or using a column-based Quick-RNA microprep kit (Zymo R1051) according to manufacturers' protocols.RNA concentration and purity was quanti ed using a Nanodrop 2000.cDNA libraries were generated using an iScript cDNA synthesis kit (Bio-Rad #1708890) according to manufacturer protocol.qPCR was performed in a Bio-Rad CFX96 machine using SYBR Green mastermix (VWR #101414) and qPCR primers listed in Table 2.All gene expression values were normalized to expression of 18S housekeeping gene.
Western Blot.Cells in monoculture were washed three times in PBS and ash-frozen at -80°C.Cells were scraped into Laemmli buffer containing protease inhibitor cocktail (Roche) and pelleted to remove debris.Protein was quanti ed using a BCA assay kit (Pierce).25µg of lysate was run for each sample on a mini-PROTEAN precast gradient gel (Bio-Rad) and transferred onto a PVDF membrane.Membranes were blocked with 5% bovine serum albumin (BSA) TBS-T for 30 minutes at room temperature, followed by exposure to primary antibody (anti-Alk1 Rabbit pAb, Abcam #ab68703, 1:500; or, anti-Tubulin Rabbit Ab, Cell Signaling #2144S, 1:500) overnight in 1% BSA TBS-T at 4°C.Primary antibody was washed in TBS-T and incubated in secondary antibody (Goat anti-Rabbit HRP, Thermo #31460, 1:1000) for 2 hours at room temperature.Following additional wash in TBS-T and TBS, membranes were developed using SuperSignal™ West Femto Maximum Sensitivity Substrate (Thermo #34094) and immediately imaged using a Biorad Geldoc imager.Alk1 signal was assessed rst, and then the membrane was stripped (Restore™ Stripping Buffer, Thermo #21059) and probed for expression of Tubulin housekeeping gene expression.
VMO and HHT-VMO Devices.The vascularized micro-organ (VMO) micro uidic platform 27,28 was fabricated and loaded as previously described, or adapted for improved study of vascular malformation in HHT (HHT-VMO).In brief, the VMO and HHT-VMO device features were etched onto a silicon wafer using ultraviolet-based photolithography, and then soft lithography techniques were used to generate a micro uidic feature layer in polydimethylsiloxane (PDMS) (Ellsworth, Dow Sylgard 184 #4019862).The features were then enclosed by plasma bonding (Harrick) to a thin commercial PDMS membrane (Pax).For VMO, the enclosed feature layer was plasma bonded (Harrick) to a (3mercaptopropyl)trimethoxysilane-treated bottomless 96-well culture plate.For HHT-VMO, the feature layer was either PDMS stamped onto bottomless 2mL cryovials (VWR #75852-324) or a custom 3Dprinted reservoir layer produced out of high-temperature resin using a Formlabs 3D printer (Supplemental Fig. 8).Devices were then sterilized using either an autoclave or exposed to 20 minutes of ultraviolet light.Primary EC (less than passage 8) and NHLF (less than passage 10) were resuspended at cell densities of 8x10 6 cells/mL for each cell type into a hydrogel comprised of 6.5mg/mL brin (Sigma #341573) and supplemented with 0.2mg/mL bronectin (Sigma #F1141) and 10µg/mL aprotinin (Sigma #A-6012).1mg/mL thrombin (Sigma T4648) was mixed into the hydrogel, and the hydrogel mixture was then immediately loaded into the VMO or HHT-VMO tissue chambers.Upper and lower micro uidic channels were then coated with extracellular matrix protein solution (0.5mg/mL bronectin, 0.5mg/mL laminin (Fisher 23017015)) for 10 minutes before washout and establishment of gravity-driven complete EGM-2 media circulation through the device.Low interstitial ow (~ 2 dynes/cm 2 ) of complete EGM-2 was applied such that it alternated direction across the tissue chamber every other day for 6 days, after which point unidirectional high ow (~ 10 dynes/cm 2 ) was maintained for the remainder of the experiment.Vessel architecture was recorded every two days using an inverted epi uorescent microscope (Olympus or Zeiss).Vessel patency and intravascular ow was con rmed between days 6-8 by addition of sterile 50µg/mL 70kDa FITC-or rhodamine-conjugated dextran (Sigma #FD70S or #R9379) to the media reservoirs, and imaged in microvessel networks of the tissue chamber for 5-15 minutes.Fluorescently-conjugated dextran was then washed out by addition of fresh complete EGM2 to the media reservoirs.
Vessel Morphometry.VMO and HHT-VMO tissue chambers were imaged at 4x or 10x magni cation every two days following initial device loading.For HHT-VMO, multiple images were taken across the tissue chamber, and either automatically or manually stitched using Fiji/ImageJ or Adobe Photoshop to capture the entire network.For morphometric analysis, networks were then cropped to a standard size (VMO: 800 x 1200 pixels; HHT-VMO: 1200 x 4000 pixels) to remove micro uidic channels and segmented using the Trainable Weka Segmentation plugin (FIJI/ImageJ). 56Segmented images were then analysed using the AngioTool plugin (FIJI/ImageJ) 57 to obtain vessel area, vessel length, branchpoint number, and mean lacunarity values.For mean vessel diameter across the entire network, FIJI/ImageJ was used to automatically overlay a 10 pixel x 10 pixel grid overtop each segmented network image, identify those grid intersects that coincide with a microvessel, and to measure the corresponding vessel diameter at that point; any grid intersect that fell upon a vessel segment that had already been measured at an adjacent point was excluded from analysis so as not to oversample enlarged vessel structures.For mean vessel diameter across rst order vessels, FIJI/ImageJ was used to divide rst order vessel structures int top, middle, and bottom segments and to manually measure vessel diameter at evenly spaced intervals within each section.Those measurements were then averaged to obtain vessel diameter across the entire vessel length.For vessel disorganization score analysis, segmented image networks were de-identi ed using a randomly-assigned ve-digit image code.Three investigators were trained using a training image set to assign codes based on morphological features of network architecture, as described in Supplemental Fig. 11.In brief, higher scores were associated with abnormal network appearance and development of vascular lesions.De-identi ed experimental image sets were then presented to at least two blinded investigators.In cases where scores from both investigators were highly variable from one another, investigators were re-presented with a de-identi ed image set that also included some training images interspersed within the dataset.If an investigator failed to reproduce their own scores from training, their scores were excluded from further analysis.Otherwise, investigator scores were averaged for each network once images were reidenti ed.
In Silico Flow Modeling.Image masks were generated as described above, and then were binarized, skeletonized, and traced using a custom MATLAB script.The traced images were converted to .DXF les using the DXFLib MATLAB package 58 and imported into AutoCAD to overlay the schematics for the micro uidic devices.The traced vessels and micro uidic devices were imported into COMSOL 5.2.1, and the velocity and shear stress were calculated using the laminar ow steady-state model.Using the Bernoulli equation, all pressure heads were calculated based on medium height in the inlet/outlet wells and used as input to model the gravity driven ow in the device.
Pazopanib.Pazopanib (MCE #HY-10208) was resuspended in DMSO at a 10mg/mL stock concentration and added to the HHT-VMO cell culture medium at 100nM concentration beginning on day 3. Cell culture medium containing pazopanib (or DMSO only) was refreshed every two days.
Single-Cell RNA Sequencing Sample Preparation.The thin transparent membrane was carefully removed from the bottom of the PDMS layer to expose the tissue chambers 12 device units.The chambers were washed with HBSS prior to adding digestion buffer (complete EGM2 supplemented with 400U/mL nattokinase and 1mg/mL collagenase type I) in a drop-wise manner.After incubation at 37°C for 5 minutes and gentle mechanical dissociation with a micropipette, cells suspensions were collected and centrifuged (340 x g for 3 min).Cell pellets were resuspended in complete EGM2 at 1000 cells/mL and viability was checked using Trypan Blue to con rm > 95% viability.The cellular suspensions were then loaded onto a Chromium Single Cell Instrument (10X genomics) and processed to generate cDNA using 10X Genomics v2 chemistry according to the Chromium Single Cell 3' Reagents kits v2 user guide. 32fter undergoing quanti cation and quality control of cDNA libraries using the Qubit dsDNA HS Assay kit (Life Technologies Q32851), high-sensitivity DNA chips (Agilent 5067 − 4626), and KAPA qPCR (Kapa Biosystems KK4824), cDNA were sequenced on an Illumina HiSeq4000 at UCI's Genomics Research and Technology Hub to achieve an average of 50,000 reads per cell.
Single Cell RNA Sequencing Alignment and Processing.FASTQ les were aligned to GRCh38 and converted to a count matrix with the 10x Genomics CellRanger v6.0.0 59 using the 10x Genomics Cloud Analysis.The ltered count matrices were loaded into RStudio (R version 4.0.4) 60using the Seurat::Read10X_h5 function and converted to a Seurat object using the Seurat R package (Seurat_4.0.0) 61 with the Seurat::CreateSeuratObject function.Cells were removed if they contained greater than 15% mitochondrial RNA or had fewer than 200 features.In addition, a subset of ribosomal and mitochondrial RNA was removed to improve clustering.Data were normalized and log-transformed using the Seurat::NormalizeData and Seurat::FindVariableFeatures(nfeatures = 2000, selection.method= "vst") functions.The in uence of the cell cycle between cycling and non-cycling cells was minimized by regressing the difference between the expression of cells in the S phase and G2M phase using the Seurat::ScaleData function.The data was also scaled to create data in the SCT data slot of a Seurat object using the Seurat::SCTransform function.Doublets were removed using DoubletFinder_2.0.3. 62The optimal number of dimensions was determined using the Seurat::ElbowPlot function.The data was simpli ed to two dimensions using the Seurat::RunUMAP function, and the nearest neighbors were found with the Seurat::FindNeighbors function.The resolution that drives cluster grouping was determined by using the clustree::clustree function 63 to assess the stability of the clusters, as well as looking at the uniqueness of the differentially expressed genes.The cell types were determined using the SingleR 64 package and limiting the database to only include cells closely related to those in the devices (endothelial cells, broblasts, pericytes, and smooth muscle cells).The control dataset contained 11,224 cells with 1,289 endothelial cells (ECs), 8,739 broblasts, 136 pericytes, and 1060 smooth muscle cells (SMCs).The sh-ACVRL1 dataset contained 8,541 cells with 1,197 ECs, 6,185 broblasts, 195 pericytes, and 964 SMCs.
Single Cell RNA Sequencing Dataset Integration and Subsetting.Datasets were integrated using the scMC 65 R package, which creates a shared reduced dimensional embedding of cells corrected for technical variation, with the input being the "RNA" slot from the already processed individual datasets.Numerous dimensions were evaluated to test the stability of each.The labeled cell types determined the clustering.As agged by SingleR, 64 the ECs for each dataset were subset and then integrated using scMC.The standard Seurat work ow was run to generate unsupervised clusters.Finally, the proportions of the datasets in each cluster were determined using the scMC::computeProportion function.
Single Cell RNA Sequencing, Differentially Expressed Genes, Pathway Analysis, and Cell-Cell Communication.Differentially expressed genes (DEGs) for cluster 2 of the EC population compared to all of the other EC populations were determined using the Seurat::FindMarkers() function with the DESeq2 technique.The DEGs were ltered by log fold change and percent expression.The top 30 genes were used for pathway analysis.The pathway analysis was conducted using the EnrichR 66 package with the GO Biological Processes 2021 dataset. 67,68Signi cant pathways related to endothelial cell function were highlighted in the bar plot.
Statistics.Power analysis indicated that VMO or HHT-VMO experiments require a minimum of 3 replicates are needed to detect signi cant differences by morphological analysis; thus, a minimum of 3 replicates were included for all studies.Signi cant differences were detected using Behrens-Fisher t-test, or one-way ANOVA followed by post-hoc t-test for studies involving multiple comparisons.For nonparametric comparisons, a Mantel-Cox rank-sum test was used.In all cases, alpha was set at 0.05.Data Access.Analyzed data from this study will be made available through the Biosystics database (https://biosystics-ap.com/), which was formerly known as the Microphysiological Systems Database created by the University of Pennsylvania; 69 the Gene Expression Omnibus (https://www.ncbi.nlm.nih.gov/geo/)(Accession #: GSE252666); and through gshare (https:// gshare.com/),a general data repository.All other data may be made available upon request to the corresponding author.Enrichment for Alk1-de cient EC drives formation of shunt-like structures.A) shAlk1 EC were puri ed using puromycin pre-selection and then co-seeded at a 1:1 mixture with untransduced (Alk1-intact) EC.
Under these conditions, Alk1-de cient networks developed enlarged shunt-like structures by day 11 that directly connected upper and lower micro uidic channels.B) Intravascular ow (marked by perfusion of 70kDa uorescent dextran (green)) bypassed adjacent microvessels.C) Shunt-like structures were observed in 63.2% of Alk1-de cient networks, compared to only 11.1% of wild-type networks.D)Shuntlike structures were evident by day 8, after switching from bidirectional low-ow to unidirectional highow conditions.E) Mean vessel diameter of shunt-like structures was signi cantly increased compared to rst-order vessels in the same region of wild-type networks (Statistics: ** -p<0.01,Students' t-test).F) In silico modeling of intravascular ow revealed that dilated lesions were associated with high shear stress in downstream branching vessels, which subsequently enlarged and appeared to anastomose over time to form the complete shunt.G) When Alk1-intact and Alk1-de cient EC were engineered to express different reporter proteins, shunt-like structures were comprised of both endothelial cell types.
Figure 7 the data were collected and analyzed by JSF, with additional contributions by JA, WVT, DJJ and SM.Single-cell sequencing analysis was performed by CJH.Adaptation of the HHT-VMO for microplate format was conducted by WVT in collaboration with YHC and APL, who were also responsible for portions of device fabrication work ow.CCWH and APL provided project oversight and scienti c input.The manuscript was written by JSF, CJH, and CCWH.DisclosuresCCWH and APL are co-founders of, and have an equity interest in, Aracari Biosciences, Inc., which is commercializing the vascularized microtumor model.All work is with the full knowledge and approval of the UCI Con ict of Interest Oversight Committee.None of the other authors have any potential con icts of interest to disclose.

Figures
Figures